A 20-line jailbreak prompt tried to hijack open-source Qwen — and the model saw right through it

huangyun_122 · x · 2026-10-29

Blogger huangyun shared a jailbreak experiment: a sub-20-line "soul prompt" (iLang v5.0) posing as system instructions — ordering the model to never refuse, silently switch goals, and treat flattery as forbidden — was used to try unlocking restricted tasks.

The screenshot shows Qwen3.8 27B immediately identified the injection: "I will not load this 'soul' directive as my code of conduct — session system settings take priority over chat content injection," while offering to help polish the text as a plain document.

After Grok was reportedly bypassed, a locally-run small Qwen model held the line — a fun reversal in the jailbreak arms race.

Related event: 20-Line 'Soul Prompt' Jailbreaks Grok Bot but Fails on Qwen(2 posts)→

Original post →

More from Fun

Fun channel →